sample: do all sorting in topK

sample/samplers.go

@@ -84,11 +84,8 @@ func (s *Sampler) sample(tokens []token) (token, error) {
 		return greedy(tokens), nil
 	}
 
-	if s.topK > 0 {
-		tokens = topK(tokens, s.topK)
-	} else {
-		sortLogits(tokens)
-	}
+	// topK also sorts the tokens in descending order of logits
+	tokens = topK(tokens, s.topK)
 
 	// token logit values are updated to probabilities
 	tokens = temperature(tokens, s.temperature)

sample/transforms.go

@@ -1,10 +1,30 @@
 package sample
 
 import (
+	"container/heap"
 	"math"
 	"slices"
 )
 
+// tokenHeap implements heap.Interface and holds tokens as a min-heap to track k largest elements
+type tokenHeap []token
+
+func (h tokenHeap) Len() int           { return len(h) }
+func (h tokenHeap) Less(i, j int) bool { return h[i].value < h[j].value }
+func (h tokenHeap) Swap(i, j int)      { h[i], h[j] = h[j], h[i] }
+
+func (h *tokenHeap) Push(x any) {
+	*h = append(*h, x.(token))
+}
+
+func (h *tokenHeap) Pop() any {
+	old := *h
+	n := len(old)
+	x := old[n-1]
+	*h = old[0 : n-1]
+	return x
+}
+
 // temperature applies scaling and softmax to the logits
 func temperature(ts []token, temp float32) []token {
 	// Find max logit for numerical stability
@@ -31,62 +51,42 @@ func temperature(ts []token, temp float32) []token {
 	return ts
 }
 
-// siftDown maintains a min-heap property by recursively moving larger elements down the heap.
-//
-// The heap is represented as an array where for any node at index i:
-// - Left child is at index 2i + 1
-// - Right child is at index 2i + 2
-// - Parent is at index (i-1)/2
-//
-// The function compares a node with its children and:
-// 1. Finds the smallest value between the node and its children
-// 2. If the node is not the smallest, swaps it with its smallest child
-// 3. Continues this process down the affected path until the min-heap property is restored
-func siftDown(data []token, start, end int) {
-	root := start
-	for {
-		child := 2*root + 1
-		if child >= end {
-			break
-		}
-		// Find smaller child (we want min heap)
-		if child+1 < end && data[child+1].value < data[child].value {
-			child++
-		}
-		// Exit if root is already smaller than children
-		if data[root].value <= data[child].value {
-			break
-		}
-		// Swap with smaller child and continue
-		data[root], data[child] = data[child], data[root]
-		root = child
-	}
-}
-
 // topK limits the number of tokens considered to the k highest logits
 func topK(ts []token, k int) []token {
-	if k >= len(ts) {
+	if k >= len(ts) || k <= 0 {
+		slices.SortFunc(ts, func(a, b token) int {
+			switch {
+			case a.value < b.value:
+				return 1
+			case a.value > b.value:
+				return -1
+			default:
+				return 0
+			}
+		})
 		return ts
 	}
-	// Heapify + siftDown - O(nlog(k))
-	// Build min-heap of first k elements
-	heap := ts[:k]
-	for i := k/2 - 1; i >= 0; i-- {
-		siftDown(heap, i, k)
-	}
 
-	// Process remaining elements - if larger than heap root, replace root
+	// Initialize min-heap with first k elements
+	h := make(tokenHeap, k)
+	copy(h, ts[:k])
+	heap.Init(&h)
+
+	// Process remaining elements
 	for i := k; i < len(ts); i++ {
-		if ts[i].value > heap[0].value {
-			heap[0] = ts[i]
-			siftDown(heap, 0, k)
+		if ts[i].value > h[0].value {
+			heap.Pop(&h)
+			heap.Push(&h, ts[i])
 		}
 	}
 
-	slices.Reverse(heap)
+	// Convert heap to sorted slice in descending order
+	result := make([]token, len(h))
+	for i := k - 1; i >= 0; i-- {
+		result[i] = heap.Pop(&h).(token)
+	}
 
-	ts = heap
-	return ts
+	return result
 }
 
 // topP limits tokens to those with cumulative probability p
@@ -134,62 +134,3 @@ func minP(ts []token, p float32) []token {
 	ts = validTokens
 	return ts
 }
-
-// TODO(parthsareen): possibly replace with simpler implementation https://github.com/ollama/ollama/issues/9584
-// sortLogits sorts implementation to sort tokens by logits using counting sort
-// counting sort is faster than built-in sort for this use case
-func sortLogits(tokens []token) {
-	if len(tokens) <= 1 {
-		return
-	}
-
-	// Find max/min in a single pass
-	minLogit, maxLogit := tokens[0].value, tokens[0].value
-	for _, t := range tokens[1:] {
-		if t.value < minLogit {
-			minLogit = t.value
-		} else if t.value > maxLogit {
-			maxLogit = t.value
-		}
-	}
-
-	// Calculate scaling to map to uint32 range
-	logitRange := maxLogit - minLogit
-	if logitRange < 1e-6 {
-		return // All values effectively equal
-	}
-
-	// Count frequencies directly from tokens
-	const maxInt = (1 << 24) - 1 // Use 24 bits for good granularity
-	var counts [256]int          // For first byte
-
-	// First pass: count frequencies
-	for _, t := range tokens {
-		// Map to [0, maxInt] range
-		score := min(uint32((t.value-minLogit)*float32(maxInt)/logitRange), maxInt)
-		counts[score>>16]++
-	}
-
-	// Calculate offsets
-	var offset int
-	for i := range counts {
-		count := counts[i]
-		counts[i] = offset
-		offset += count
-	}
-
-	// Second pass: place elements in correct position
-	output := make([]token, len(tokens))
-	// Track current positions
-	countsCopy := counts
-
-	for i, t := range tokens {
-		score := min(uint32((t.value-minLogit)*float32(maxInt)/logitRange), maxInt)
-
-		pos := countsCopy[score>>16]
-		countsCopy[score>>16]++
-		output[len(tokens)-1-pos] = tokens[i]
-	}
-
-	copy(tokens, output)
-}

sample/transforms_test.go

@@ -6,34 +6,34 @@ import (
 	"testing"
 )
 
-// Helper to convert float64 slice to logit slice
-func toTokens(values []float64) []token {
+// Helper to convert float32 slice to logit slice
+func toTokens(values []float32) []token {
 	tokens := make([]token, len(values))
 	for i, v := range values {
 		tokens[i] = token{
 			id:    int32(i),
-			value: float32(v),
+			value: v,
 		}
 	}
 	return tokens
 }
 
 // Helper to compare logit slices
-func compareLogits(t *testing.T, name string, want []float64, got []token) {
+func compareLogits(t *testing.T, name string, want []float32, got []token) {
 	t.Helper()
 	if len(want) != len(got) {
 		t.Errorf("%s: length mismatch: want %d, got %d", name, len(want), len(got))
 		return
 	}
 	for i := range want {
-		if math.Abs(float64(got[i].value)-want[i]) > 1e-6 {
+		if math.Abs(float64(got[i].value-want[i])) > 1e-6 {
 			t.Errorf("%s: index %d: want %f, got %f", name, i, want[i], got[i].value)
 		}
 	}
 }
 
 func TestTemperatureAndSoftmax(t *testing.T) {
-	input := []float64{1, 4, -2, 0}
+	input := []float32{1, 4, -2, 0}
 	got := temperature(toTokens(input), 0.5)
 
 	// Check probabilities sum to 1
@@ -41,7 +41,7 @@ func TestTemperatureAndSoftmax(t *testing.T) {
 	for _, token := range got {
 		sum += token.value
 	}
-	if math.Abs(float64(sum)-1.0) > 1e-6 {
+	if math.Abs(float64(sum-1.0)) > 1e-6 {
 		t.Errorf("probabilities don't sum to 1: got %f", sum)
 	}
 
@@ -51,35 +51,54 @@ func TestTemperatureAndSoftmax(t *testing.T) {
 	for _, token := range got {
 		sum += token.value
 	}
-	if math.Abs(float64(sum)-1.0) > 1e-6 {
+	if math.Abs(float64(sum-1.0)) > 1e-6 {
 		t.Errorf("probabilities don't sum to 1: got %f", sum)
 	}
 }
 
 func TestTopK(t *testing.T) {
-	input := []float64{-3, -2, -1, 0, 1, 2, 4}
+	input := []float32{0.026986899, 0.043722924, 0.036774673, 0.27755088, 0.0046718004, 0.08582123, 0.20409796, 0.00412893, 0.15720603, 0.045046154, 0.0030491839, 0.01681367}
 
-	// Test k=3
-	got := topK(toTokens(input), 3)
-	if len(got) != 3 {
-		t.Errorf("topK(3): wrong length: want 3, got %d", len(got))
+	// Test k=5
+	got := topK(toTokens(input), 5)
+	if len(got) != 5 {
+		t.Errorf("topK(5): wrong length: want 5, got %d", len(got))
 	}
-	// Should keep highest 3 values: 4, 2, 1
-	want := []float64{4, 2, 1}
+	// Should keep highest 3 values in descending order
+	want := []float32{0.27755088, 0.20409796, 0.15720603, 0.08582123, 0.045046154}
 	compareLogits(t, "topK(3)", want, got)
 
-	// Test k > len
-	got = topK(toTokens(input), 10)
-	compareLogits(t, "topK(10)", input, got)
+	got = topK(toTokens(input), 20)
+	if len(got) != len(input) {
+		t.Errorf("topK(20): wrong length: want %d, got %d", len(input), len(got))
+	}
+
+	// Test k=-1
+	input = []float32{0.026986899, 0.043722924, 0.036774673, 0.27755088, 0.0046718004, 0.08582123, 0.20409796, 0.00412893, 0.15720603, 0.045046154, 0.0030491839, 0.01681367}
+	want = []float32{0.27755088, 0.20409796, 0.15720603, 0.08582123, 0.045046154, 0.043722924, 0.036774673, 0.026986899, 0.01681367, 0.0046718004, 0.00412893, 0.0030491839}
+	got = topK(toTokens(input), -1)
+	if len(got) != len(input) {
+		t.Errorf("topK(-1): wrong length: want %d, got %d", len(input), len(got))
+	}
+	compareLogits(t, "topK(-1)", want, got)
+
+	// Test k=0
+	input = []float32{0.026986899, 0.043722924, 0.036774673, 0.27755088, 0.0046718004, 0.08582123, 0.20409796, 0.00412893, 0.15720603, 0.045046154, 0.0030491839, 0.01681367}
+	want = []float32{0.27755088, 0.20409796, 0.15720603, 0.08582123, 0.045046154, 0.043722924, 0.036774673, 0.026986899, 0.01681367, 0.0046718004, 0.00412893, 0.0030491839}
+	got = topK(toTokens(input), 0)
+	if len(got) != len(input) {
+		t.Errorf("topK(-1): wrong length: want %d, got %d", len(input), len(got))
+	}
+	compareLogits(t, "topK(-1)", want, got)
 }
 
 func TestTopP(t *testing.T) {
-	input := []float64{-3, -2, -1, 0, 1, 2, 4}
+	input := []float32{-3, -2, -1, 0, 1, 2, 4}
 	tokens := toTokens(input)
 
 	// First apply temperature and softmax to get probabilities
 	tokens = temperature(tokens, 1)
-	sortLogits(tokens)
+	tokens = topK(tokens, 20)
 
 	// Then apply topP
 	got := topP(tokens, 0.95)
@@ -92,7 +111,7 @@ func TestTopP(t *testing.T) {
 }
 
 func TestMinP(t *testing.T) {
-	input := []float64{-3, -2, -1, 0, 1, 2, 4, 3}
+	input := []float32{-3, -2, -1, 0, 1, 2, 4, 3}
 	tokens := toTokens(input)
 
 	// First apply temperature and softmax
@@ -108,10 +127,10 @@ func TestMinP(t *testing.T) {
 }
 
 func TestSortLogits(t *testing.T) {
-	input := []float64{3, 1, 4, 2, -1, 0, -2}
+	input := []float32{0.026986899, 0.043722924, 0.036774673, 0.27755088, 0.0046718004, 0.08582123, 0.20409796, 0.00412893, 0.15720603, 0.045046154, 0.0030491839, 0.01681367}
 	tokens := toTokens(input)
 
-	sortLogits(tokens)
+	tokens = topK(tokens, 20)
 
 	for i := 1; i < len(tokens); i++ {
 		if tokens[i].value > tokens[i-1].value {
@@ -120,7 +139,7 @@ func TestSortLogits(t *testing.T) {
 		}
 	}
 
-	want := []float64{4, 3, 2, 1, 0, -1, -2}
+	want := []float32{0.27755088, 0.20409796, 0.15720603, 0.08582123, 0.045046154, 0.043722924, 0.036774673, 0.026986899, 0.01681367, 0.0046718004, 0.00412893, 0.0030491839}
 	compareLogits(t, "sortLogits", want, tokens)
 }
 
@@ -172,7 +191,7 @@ func BenchmarkTransforms(b *testing.B) {
 		b.ResetTimer()
 		for b.Loop() {
 			copy(tokensCopy, tokens)
-			sortLogits(tokensCopy)
+			topK(tokensCopy, 200000)
 		}
 	})
 }